Merge pull request #40925 from BerriAI/litellm_proxy_error_headers_from_litellm_response_headers

fix(proxy): forward provider request id headers on mapped error responses

litellm/proxy/common_request_processing.py

@@ -3452,15 +3452,13 @@ class ProxyBaseLLMRequestProcessing:
             # a failed request reports no timing, matching /v1/chat/completions
             read_timing_from_logging_obj=False,
         )
-        # Extract headers from exception - check both e.headers and e.response.headers
         headers = getattr(e, "headers", None) or {}
         if not headers:
-            # Try to get headers from e.response.headers (httpx.Response)
             _response: Final = attribute_of(e, "response")
-            if _response is not None:
-                _response_headers: Final = getattr(_response, "headers", None)
-                if _response_headers:
-                    headers = get_response_headers(dict(_response_headers))
+            _response_headers: Final = getattr(_response, "headers", None) if _response is not None else None
+            _provider_headers: Final = _response_headers or getattr(e, "litellm_response_headers", None)
+            if _provider_headers:
+                headers = get_response_headers(dict(_provider_headers))
         headers.update(custom_headers)
 
         # Call response headers hook for failure

tests/test_litellm/proxy/test_common_request_processing.py

@@ -8400,6 +8400,41 @@ async def test_handle_llm_api_exception_forwards_provider_headers_on_http_status
     assert exc_info.value.headers["llm_provider-x-amzn-requestid"] == "req-passthrough-500"
 
 
+@pytest.mark.asyncio
+async def test_handle_llm_api_exception_forwards_litellm_response_headers_when_response_is_synthetic():
+    """Exception mapping hands the proxy a mapped error whose ``response`` is a synthetic empty
+    ``httpx.Response`` and parks the provider's real headers on ``litellm_response_headers``.
+    The client must still get the provider request id, as it does on a 200.
+    """
+    import httpx
+
+    from litellm.proxy._types import ProxyException, UserAPIKeyAuth
+
+    mapped = litellm.BadRequestError(
+        message="OpenAIException - max_tokens is too large: 999999999.",
+        model="gpt-4o-mini",
+        llm_provider="openai",
+    )
+    mapped.litellm_response_headers = httpx.Headers({"x-request-id": "req_openai_400"})
+    assert dict(mapped.response.headers) == {}
+
+    processor = ProxyBaseLLMRequestProcessing(data={})
+    proxy_logging_obj = MagicMock()
+    proxy_logging_obj.post_call_failure_hook = AsyncMock(return_value=None)
+    proxy_logging_obj.post_call_response_headers_hook = AsyncMock(return_value={})
+
+    with pytest.raises(ProxyException) as exc_info:
+        await processor._handle_llm_api_exception(
+            e=mapped,
+            user_api_key_dict=UserAPIKeyAuth(api_key="sk-test"),
+            proxy_logging_obj=proxy_logging_obj,
+        )
+
+    assert exc_info.value.code == "400"
+    assert "max_tokens is too large: 999999999." in exc_info.value.message
+    assert exc_info.value.headers["llm_provider-x-request-id"] == "req_openai_400"
+
+
 class TestBackgroundResponseRetrievalGovernance:
     """LIT-7175: retrieving a background Response attaches the model's post_call policy pipelines."""